PD：banana imagesize by ratio

PD:banana imagesize by ratio (PDbananaImagesizeByRatio) resizes an image to one of eight preset size-and-ratio buckets - 1:1 at 1024, 4:3 at 1184×864, 16:9 at 1344×768, and so on - and decides what to do with the mismatch: crop, pad, or stretch. It's a one-node answer to "make this image exactly this aspect ratio at exactly this resolution," which in 2026 is a surprisingly recurring request because those presets are exactly the size buckets the SDXL-era latent models were trained on.

How it works

You pick a preset_size, and the node scales the image to fit. The interesting part is resize_mode, which determines how it handles the aspect mismatch:

• crop - scales up until the image covers the target, then crops the overflow. Nothing lost is nothing... well, content is lost at the edges.
• pad (default) - scales to fit inside the target and fills the letterbox with the padding_color.
• stretch - just force-resizes, aspect ratio be damned. Almost always the wrong choice unless the distortion is acceptable.

image_location (top/down/left/right/center) picks where the image sits in the crop window or the padded canvas. padding_color is where it gets fun: black, white, or noise. The noise option is a genuinely thoughtful touch for training data - a noisy pad is less likely to be learned as a feature than a flat black bar. All scaling is Lanczos, and the output is a clean RGB tensor at exactly the target dimensions.

The inputs that matter

• preset_size - the bucket. The dropdown carries both ratio and pixels, so what you see is what you get.
• resize_mode - crop vs pad vs stretch. This is the decision that shapes your output.
• image_location - where content sits after the transform.
• padding_color - black/white/noise. Pick noise for training sets, black/white for clean presentation.

Installing

Part of the 7BEII/Comfyui_PDuse pack. ComfyUI Manager → search "Comfyui_PDuse" → install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/7BEII/Comfyui_PDuse
cd Comfyui_PDuse
pip install -r requirements.txt

Restart after. No models to download; the dependencies are the pack's usual Pillow/numpy set.

Gotchas

The presets are fixed - there's no custom-size input, so if you need 768×1152 you're out of luck with this node; it's presets or nothing. And remember stretch doesn't preserve aspect ratio, which catches people who reach for it expecting a simple resize. If you're feeding a batch of images, note the node squeezes a batch dimension - it expects a single image tensor, so loop your batch through it or feed frames one at a time. For dataset prep it's the noise-pad option that makes it worth having in the toolbox.